Chubby♨️ · @kimmonismus · X·2026-09-03 04:45·1小时前
AI 导读

Mostik 称其潜空间桥接协议让前沿模型与 4B 边缘模型直接传递 hidden states,无需文本、双方均不微调;753B 模型读题、4B 模型作答时达到前沿模型约 80% 的准确率、速度快 20 倍,且比分数匹配的中型模型少用 2.5 倍算力。

Chubby♨️@kimmonismus
47AI 编辑部评分,满分 100
2026-09-03 04:45· 1小时前
AI 导读

Mostik 称其潜空间桥接协议让前沿模型与 4B 边缘模型直接传递 hidden states,无需文本、双方均不微调;753B 模型读题、4B 模型作答时达到前沿模型约 80% 的准确率、速度快 20 倍,且比分数匹配的中型模型少用 2.5 倍算力。

A 753B model reads the problem: A 4B model writes the answer.

Mostik reports that its latent bridge lets the 4B model close half the performance gap to the 753B model

The bridged pair uses 2.5x less compute than a score-matched mid-sized model

A fascinating alternative to the text-based handoffs used by most multi-model systems today. Incredible!

Sasha Malyshevameet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the c...